Offre en lien avec l’Action/le Réseau : – — –/Innovation

Laboratoire/Entreprise : SAMOVAR
Durée : 6
Contact : aikaterini.tzompanaki@cyu.fr
Date limite de publication : 2024-01-15

Contexte :
With the emergence of the Internet of Things (IoT) and computing (cloud-edge) continuum technologies
infrastructures are becoming more sensorized, fueling the development of smart space ecosystems and
improving societal quality of life. As a result, smart spaces are becoming popular in many domains,
including healthcare, education, building management, and more. This integration with the IoT brings
much potential in revolutionizing the way that these environments operate. Initially, IoT sensors mea-
sure physical phenomena (temperature, energy consumption, luminosity) in a continuous way, producing
streams of data. Such data is often used for analysis and predictions to either optimize different criteria
(e.g., occupancy, user comfort, energy consumption, etc.), or identify and anticipate problems. To man-
age and control the data generated by IoT devices, AI algorithms can be used to enable smarter, more
efficient, and more responsive devices to their environment.

Sujet :
To enable efficient decision making, it is increasingly common that predictions be accompanied by
explanations, i.e., pieces of information either on the data, the models, or both for giving insights on
why the predictions were made. For example, a smart building employs IoT devices to measure energy
consumption of different plugs in households, as well as temperature of different rooms. Using this
measurements as training data, a Machine Learning model can be used to predict the energy consumption
of the air-conditioning appliance plug (label) using different IoT data features (e.g., temperature). Since
this is a continuous setting scenario, at time t1 we may have gathered D1 data, which we use to train a
decision tree M1. At a later time t2, incrementally we have obtained D1 ⊂ D2, along with a new trained
model M2. Let’s assume the following predictions made by M1 and M2 for a test point p: y1 = M1(p) =
10watts and y2= M2(p) = 50watts. An example of traditional local explanations for these two predictions
could be e1 : p.temperature = 20 while e2 : p.temperature = 35. We argue that a more informative
and correct explanation would be the information that the distribution of the training data has changed;
it is not only that the current temperature is at 35 degrees, but also that the timestamps with high
temperatures appear more often in the data than before t1.
This internship aims to study the usage of data distribution changes through time for the construction
of more pertinent XAI models for IoT-enhanced spaces. Use cases will be provided from existing data
models and data instances from IoT devices deployed in smart spaces of the Institut Polytechnique de
Paris (IP Paris). The successful candidate will be considered for a 3 year PhD contract at the end of the internship

Profil du candidat :
Internship Objectives
The selected candidate will be working on the following tasks:
– Get familiar with data models for smart spaces.
– Leverage datasets of smart spaces for prediction and decision making.
– Study data drifts and distribution changes in datasets for IoT space predictions.
– Propose explanation formalizations based on data distribution changes.

Formation et compétences requises :
Master 2 on computer science or last year of engineering school.

Adresse d’emploi :
Telecom SudParis, Evry or Palaiseau

Document attaché : 202311091058_XAI_IoT_internship.pdf